Foundation Models for Flexible and Safe Operation of Multi-Energy Systems

Foundation Models for Flexible and Safe Operation of Multi-Energy Systems

14/09/2026 Autre financement public

Mines Paris-PSL

Foundation Models for Flexible and Safe Operation of Multi-Energy Systems

• Energie

Foundation models, Artificial intelligence, Energy digitalization, Multi-energy systems, Operational safety, Flexibility

Context and challenges:

The European Union Artificial Intelligence Act defines a Foundation Model (FM) as ‘AI model that is trained on broad data at scale, is designed for generality of output, and can be adapted to a wide range of distinctive tasks’. FMs employ generative AI approaches based on large language models or advanced encoding/decoding architectures, harvesting vast datasets autonomously via self-supervision. FMs show better performance in zero-shot and few-shot learning benchmarks compared with traditional statistical or machine learning models. Relevant developments for the energy sector include so-called Time-Series Foundation Models (TSFM) able to perform forecasting of multiple quantities or sites simultaneously thanks to large pre-trained corpus of diverse datasets. TSFMs tend to be large models, with tens to hundreds of millions of parameters, see e.g. Chronos, TimesFM, Moirai, LLMTime. Conversely, smaller FMs have been dedicated to specific predictive tasks for power systems such as forecasting of renewable power production or electricity prices (WindFM, PowerPM). Alternatives have also been explored to learn better on different resolutions and improve fine-tuning using exogenous information (see e.g. Tiny Time Mixers).

Beyond pure predictions, FM frameworks have been proposed to approximate complex optimization problems in power systems (see e.g. the ongoing GridFM initiative). Leveraging graph-based neural networks help solve optimal power flow with reduced computational cost. However, these frameworks are early-stage and remain limited to electricity-focused applications. As these approaches ignore dynamics in multi-energy operation between energy vectors (e.g. electricity / heat / gas, hydrogen), they cannot guarantee feasible and secure combined operation of electricity, heat and gas systems.

Main objective of the thesis:

This PhD explores the development of foundation models (FMs) as surrogate models for multi-energy system optimization to support operational decisions such as flexibility booking and security-constrained optimal power and energy flows.
The development will be supported by two key technical objectives: (1) construction of relevant datasets, starting with the identification and collection of adequate datasets, and complemented after with synthetically generated datasets to enhance the FM learning capabilities; (2) establish a learning approach to extract structural information in a self-supervised manner and fine-tune to downstream tasks for multi-energy systems.

Methodology and expected results:

The proposed FM will be pretrained on large datasets of operational scenarios covering electricity, heat, gas, hydrogen, storage, and conversion technologies, representative of current and future European infrastructures. These datasets will be augmented by a multi-energy flow model developed at PERSEE and through Hardware-in-the-Loop (HIL) experiments. The FM will employ physics-inspired Machine Learning to guarantee physically feasible operation. More generally trustworthy AI principles including explainability and uncertain quantification will be embedded. After pretraining, the architecture will be refined to better represent interactions among energy vectors, investigating among others mixture of experts and fine-tuning following prescriptive analytics principles. The trained FM will act as a fast, high-fidelity surrogate for predictive management and distributed or federated learning experiments, avoiding repeated full multi-energy flow computations. Integration with a HIL platform will validate predictions in realistic operational settings, enabling real-time interaction with controllers and devices.

The expected outcome is a FM-based framework that complements real-time control and predictive management by providing rapid, reliable, and generative support for scenario exploration and operational decision-making across time scales from milliseconds to days.

Collaborations:

This thesis is supported by the PEPR program FutuRE funded underFrance 2030:

The project is also linked to our participation to the European initiative AI.
Grids supported by CRESYM and the European Commission.

Mines Paris-PSL

ThePERSEE Centeris one of the 18 research centers ofMINESParis.